def get_example(self):
if self.mix: # Training phase of BC learning
# Select two training examples
while True:
sound1, label1 = self.base[random.randint(
0,
len(self.base) - 1)]
sound2, label2 = self.base[random.randint(
0,
len(self.base) - 1)]
if label1 != label2:
break
sound1 = self.preprocess(sound1)
sound2 = self.preprocess(sound2)
# Mix two examples
r = np.array(random.random())
sound = U.mix(sound1, sound2, r, self.opt.fs).astype(np.float32)
eye = np.eye(self.opt.nClasses)
label = (eye[label1] * r + eye[label2] * (1 - r)).astype(
np.float32)
else: # Training phase of standard learning or testing phase
sound, label = self.base[i]
sound = self.preprocess(sound).astype(np.float32)
label = np.array(label, dtype=np.int32)
if self.train and self.opt.strongAugment:
sound = U.random_gain(6)(sound).astype(np.float32)
return sound, label
def _generate_sample(self, fIdA, fIdB):
'''
Takes 2 audio fileidx, preprocess them, mix them (if needed)
'''
sound1, label1 = self.fId_to_sound(fIdA)
sound2, label2 = self.fId_to_sound(fIdB)
if self.n_classes == 10:
lbl_indexes = {0:0, 1:1, 10:2, 11:3, 12:4,
20:5, 21:6, 38:7, 40:8, 41:9}
label1 = lbl_indexes[label1]
label2 = lbl_indexes[label2]
if self.mix: # Mix two examples
r = np.array(random.random())
sound = U.mix(sound1, sound2, r, self.audio_rate) # delays
sound = sound.astype(np.float32)
eye = np.eye(self.n_classes)
label = (eye[label1] * r + eye[label2] * (1 - r))
label = label.astype(np.float32)
else:
sound, label = sound1, label1
if self.strongAugment:
sound = U.random_gain(6)(sound).astype(np.float32)
sound = sound[:, np.newaxis]
return sound, label
def _data_gen(self):
self.stop = False
while not self.stop:
idxs1 = list(self.df.index)
idxs2 = list(self.df.index)
if self.randomize:
random.shuffle(idxs1)
random.shuffle(idxs2)
for idx1, idx2 in zip(idxs1, idxs2):
fname1 = self.df.filename[idx1]
fname2 = self.df.filename[idx2]
sound1 = self.fname_to_wav(fname1)
sound2 = self.fname_to_wav(fname2)
sound1 = self.preprocess(sound1)
sound2 = self.preprocess(sound2)
label1 = self.df.target[idx1]
label2 = self.df.target[idx2]
if self.n_classes == 10:
lbl_indexes = [0, 1, 10, 11, 12, 20, 21, 38, 40, 41]
label1 = lbl_indexes.index(label1)
label2 = lbl_indexes.index(label2)
if self.mix:
# Mix two examples
r = np.array(random.random())
sound = U.mix(sound1, sound2, r, self.audio_rate)
sound = sound.astype(np.float32)
eye = np.eye(self.n_classes)
label = (eye[label1] * r + eye[label2] * (1 - r))
label = label.astype(np.float32)
else:
sound, label = sound1, label1
if self.strongAugment:
sound = U.random_gain(6)(sound).astype(np.float32)
sound = sound[:, np.newaxis]
yield sound, label